Fast Adversarial Attacks with Gradient Prediction

Kamil Ciosek, Aleksandr V. Petrov, Nicolò Felicioni, Konstantina Palla · 2026

Read Fast Adversarial Attacks with Gradient Prediction by Kamil Ciosek, Aleksandr V. Petrov, Nicolò Felicioni, Konstantina Palla (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-14868.txt.

Description

Ciosek, Petrov, Felicioni and Palla introduce fast adversarial attacks with gradient prediction. Generating adversarial examples is core to robustness evaluation and red-teaming, yet even fast attacks like FGSM are limited by the backward pass. The authors eliminate the backward pass by forecasting input gradients from forward-pass hidden states via linear regression. Exact affine conditional gradient means are derived, showing optimality in the idealised NTK setting. Much of FGSM performance is recovered at a fraction of the time, a claimed five hundred and thirty-two percent increase.

Catalog record

CREATORKamil Ciosek, Aleksandr V. Petrov, Nicolò Felicioni, Konstantina Palla
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56310
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.14868

Browse the full library · About the Tower of Babel